Researchers at EPFL’s Laboratory of Computational Science and Modeling (COSMO) have reached a significant milestone in material science, reaching the top position on Matbench Discovery, the leading benchmarking platform for machine-learning interatomic potentials.
Beyond setting a new performance reference, the result highlights a thoughtful approach aimed at making machine learning more reliable and useful for scientific research. The Matbench Discovery leaderboard has long been dominated by models developed at Meta, which can deploy huge computational resources to train them. The achievement of reaching the top position was made possible by PET-OAM-XL, a new model, which builds on the foundations laid by the recently published PET-MAD model.
Matbench Discovery: A shared benchmark for materials discovery
For a few years, Matbench Discovery has become a central reference point in materials science, comparable to ImageNet in computer vision. It provides a shared, trusted framework for evaluating how well machine-learning models perform on realistic materials discovery tasks, such as predicting whether a crystal structure is stable, how it relaxes, and how efficiently a model can guide large-scale screening efforts. Because all models are evaluated on the same data and under the same protocol, Matbench Discovery has become essential for both developers and end-users across the community, and models that are not represented there often struggle to gain broader adoption.
The model behind this result, PET-OAM-XL, is the latest evolution of a line of research that began with PET-MAD, whose foundational paper was recently published. PET-MAD was conceived as a lightweight, general-purpose model designed for advanced materials simulations, particularly molecular dynamics and other scenarios where robustness and scalability matter. PET-OAM-XL builds on this foundation by scaling the architecture and training it on datasets more specifically tailored to materials discovery tasks.
In practical terms, PET-OAM-XL is a larger model trained with a more specific focus. The aim was to go beyond PET-MAD by concentrating on the kinds of tasks emphasized in Matbench Discovery (e.g., stability prediction, crystal structures, and related properties) while preserving the generality of the underlying approach. Whereas PET-MAD was designed as a global model for atomic-scale simulations, PET-OAM-XL represents a more focused version, optimized for discovery without sacrificing the versatility that defined the original model.
Rethinking model design in atomistic machine learning
At the heart of PET-OAM-XL lies a distinctive architectural choice. Many state-of-the-art models in atomistic machine learning explicitly encode physical symmetries, such as how atomic systems behave under rotations or inversions, directly into their architecture. This is a powerful and well-established approach. PET-OAM-XL takes a different path. Rather than hard-coding these constraints in the architecture, the model is designed to learn symmetries from data, using extensive data augmentation.
This approach mirrors how modern vision models learn invariances in classification tasks. A cat remains a cat whether an image is rotated, flipped, or slightly recolored. Rather than enforcing such invariances by construction, models can be exposed to many transformed examples through data augmentation and learn what truly matters. PET-OAM-XL follows the same principle. In atomic systems, the underlying interactions are invariant to rotations and orientations of the structure, and instead of hard-coding these symmetries into the architecture or loss function, the model learns them directly from the data.
Learning symmetries rather than enforcing them challenges conventional wisdom in parts of the field. Building on the same logic illustrated by the image-classification analogy, the results show that, when combined with sufficiently diverse and carefully curated data, this approach can yield models that are both accurate and remarkably flexible. By avoiding rigid architectural constraints, PET-OAM remains closer to widely used transformer-based architectures, making it easier to scale, adapt, and integrate into existing simulation workflows.
Given the central importance of data, it might also be time to reconsider how reference datasets are built. For years, most datasets used to train interatomic potentials focused almost exclusively on stable, low-energy structures —the idealized configurations found in materials databases. While invaluable for materials discovery tasks, such data represents only a small fraction of the atomic configurations encountered in real simulations. Recognizing this limitation, researchers at the COSMO lab trained the general-purpose PET-MAD model using the MAD (Massive Atomistic Diversity) dataset which incorporates a high degree of chemical and structural diversity. This diversity enables the model to learn not only what stable materials look like, but also how atoms behave far from ideal conditions, contributing to both competitive accuracy and strong transferability.
Making models usable beyond benchmarks
PET-OAM-XL was developed not only for benchmarks performance but also with a strong emphasis on usability in real research environments. From the outset, researchers at the COSMO Lab paid close attention to the practical constraints faced by scientists running large-scale and long-timescale simulations. Rather than optimizing solely for benchmark performance, PET-OAM was engineered to scale efficiently to large systems without prohibitive memory or computational costs.
This focus on practical deployment is reflected in the model’s integration into standard atomistic simulation engines, allowing researchers to use it directly for molecular dynamics, studies of ionic conductivity, phase transitions, alloy stability, and simulations relevant to organic chemistry and experimental characterization. The accompanying open-source release, clear documentation, and straightforward installation further lower the barrier to adoption.
More details are available in the recent preprint Pushing the limits of unconstrained machine-learned interatomic potentials.
What comes next for PET models
Building on MAD as a strong conceptual foundation, characterized by diverse structures, consistent reference calculations, and a unified computational protocol, the team is already planning the next iteration of PET-MAD.
Developed within the framework of the NCCR MARVEL, the project also benefited from strong support across the Swiss research landscape and from partner organizations, including the EPFL AI Center, the SwissAI Initiative, and Swiss National Supercomputing Centre (CSCS), which were instrumental in achieving these results.
“Without their community and resources, I don’t think it would have been possible for us to compete with the giants.” says Filippo Bigi, Doctoral Assistant at EPFL COSMO Lab. For the future, the goal is to significantly expand dataset diversity, improve the quality of quantum-mechanical (DFT) reference data, and extend coverage to additional chemical elements.
Benchmarks provide an essential reference point, although real-world applications do not always mirror these conditions. In this respect, PET-based models have demonstrated strong practical relevance and, adding to their achievements, have recently reached top performance in the AIS25 AI4Materials Nanoparticles Challenge, an international competition focused on using machine-learning models to predict synthesis recipes for novel nanoparticles, where researchers from the COSMO Lab were among the winners of the first stage.
Overall, the success of PET-OAM-XL on Matbench Discovery underscores an approach grounded in both technical innovation and practical insight, demonstrating that carefully designed data, flexible architectures, and a strong emphasis on real-world usability can be just as important as raw benchmark scores in advancing the frontiers of materials discovery.
References
Bigi, F., Pegolo, P., Mazitov, A., & Ceriotti, M. (2026). Pushing the limits of unconstrained machine-learned interatomic potentials. arXiv preprint arXiv:2601.16195.
Mazitov, A., Bigi, F., Kellner, M. et al. PET-MAD as a lightweight universal interatomic potential for advanced materials modeling. Nat Commun 16, 10653 (2025).
The PET-OAM-XL model on Matbench Discovery : https://matbench-discovery.materialsproject.org/models/pet-oam-xl-1.0.0
COSMO lab GitHub repository: https://github.com/lab-cosmo/upet
Funding & Support
Swiss National Supercomputing Centre (CSCS)
NCCR Marvel
EPFL AI Center
Swiss AI Initiative
Author: Nicolas Machado